TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models

TILDE. Teacher-in-the-Loop Distillation from Embodied Foundation Models. By Yuxiang Lu and six others, from the University of Hong Kong. Published on the eighteenth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Large-scale action pretraining gives Embodied Foundation Models, or EFMs, strong manipulation capabilities, with particular strengths in generalization and in closed-loop robustness. Reproducing that pretraining is the obstacle. It requires substantial robot data and substantial computation. The question is how to carry those capabilities over to policies that never went through action pretraining themselves.

TILDE, short for Teacher-In-the-Loop Distillation from EFMs, treats the pretrained model as a teacher and draws on it in two ways at once. One is on-policy distillation at student-induced states, meaning the states the student's own behaviour leads it into. The other is learning from recovery trajectories that the teacher itself executes.

Gathering those recoveries is the job of AutoDAgger, an autonomous data aggregation pipeline. A robotic reward model monitors task progress and triggers teacher intervention, so recovery experience is collected without human monitoring and without manual control. Compared with the human-in-the-loop method, AutoDAgger speeds up recovery-trajectory collection by more than two times.

Taken together, TILDE unifies three signals: learning from demonstrations, on-policy teacher supervision, and the teacher's corrective behaviour. Across three real-world tasks it improves task performance by up to four point one times over behavior cloning. It also supports effective cross-architecture distillation, so the student need not share the teacher's design.

The arrangement lets a student learn from where it actually ends up, mistakes included, with the teacher stepping in to demonstrate the way back. Handing the decision of when to step in to a reward model is what removes the person from the loop, and the person was the slow part.